In producing oil and gas wells, two-phase flow of gas and liquid is inevitable. Gas density is lighter than liquid density, as a result, gas travels faster than liquid, leaving the liquid phase to build up in pipe segments. The amount of liquid occupying each pipe segment varies. This phenomenon is called liquid holdup, which is defined as the ratio of the liquid volume in a pipe element to the volume of the pipe element. Liquid holdup presents challenges in calculating mixture physical properties and two-phase flow pressure drop. Estimating liquid holdup is the first step to investigate production wells problems. Early detection and identification of liquid holdup (HL) in oil and gas wells is needed to reduce the maintenance downtime and thus increase production. This is significantly crucial in designing oil and gas facilities. Consequently, many studies on predicting liquid holdup have been carried out in the past, considering different flow conditions, resulting in a large number of empirical correlations with various degrees of accuracy. Machine learning approaches in predicting liquid holdup in multiphase flows have been recently studied to improve the prediction accuracy compared to the existing empirical correlations. However, these approaches ignored the heuristic feature importance of the input parameters to the predicted HL values. In our study, a machine learning predictive model, boosted decision tree regression (BDTR), is trained, tested and evaluated in predicating HL in multiphase flows in oil and gas wells. Decision trees are considered as non-parametric machine learning models. The datasets used in training and testing the predictive model are experimental and they were collected from literature (111 data-points). Air-kerosene and air-water mixtures were used in obtaining the 111 experimental data-points. Results show that, the proposed BDTR model outperforms the best empirical correlations and the fuzzy logic model used in estimating HL in gas-liquid multiphase flows. For the built model, the most important input feature in estimating HL is the superficial gas velocity (Vsg). The empirical correlations developed in the past for identifying HL in the multiphase flow phenomenon can only be applied under certain flow conditions by which they were originally developed but this machine learning model does not suffer from this limitation. To the best of our knowledge, the present study is the first work that shows how the decision forest regression predictive models can accurately predict HL. Using the BDTR model with its interpretable representation, one can clearly determine the heuristic feature importance of the input features used in building the model. Heuristic feature importance can help in having a better insight of the issues associated with the HL studies, such as the liquid loading phenomenon.